Contextual Classification withFunctional Max-Margin Markov Networks
Dan Munoz Drew BagnellNicolas Vandapel Martial Hebert
2
Problem
Wall
Vegetation
Ground
Tree trunk
Sky
Support
Vertical
Geometry Estimation(Hoiem et al.)
3-D Point Cloud Classification
Our classifications
3
Room For Improvement
4
Approach: Improving CRF Learning
Gradient descent (w) “Boosting” (h)
+ Better learn models with high-order interactions+ Efficiently handle large data & feature sets+ Enable non-linear clique potentials
• Friedman et al. 2001, Ratliff et al. 2007
5
Conditional Random FieldsLafferty et al. 2001
Pairwise model
MAP Inference
yi
xLabels
6
Parametric Linear Model
Weights
Local features that describe label
7
Associative/Potts Potentials
Labels Disagree
8
Overall Score
Overall Score for a labeling y to all nodes
9
Learning Intuition Iterate
• Classify with current CRF model
• If (misclassified)
• (Same update with edges)
φ( ) increase score
φ( ) decrease score
10
Max-Margin Structured Prediction
min Best score fromall labelings (+M)
Score withground truth labelingw
Ground truth labels
Taskar et al. 2003
Convex
11
Descending† Direction
Labels from MAP inference
Ground truth labels
(Objective)
12
Learned Model
13
Unit step-size
Update Rule
Ground truth Inferred
wt+1+= - + -
, and λ = 0
14
Iterate
• If (misclassified)
Verify Learning Intuition
wt+1 += - + -
φ( ) increase score
φ( ) decrease score
15
Alternative Update1. Create training set: D
• From the misclassified nodes & edges
, +1
, -1
, +1
, -1D =
16
Alternative Update1. Create training set: D2. Train regressor: ht
ht(∙)D
17
Alternative Update1. Create training set: D2. Train regressor: h3. Augment model:
(Before)
18
Functional M3N Summary Given features and labels for T iterations
• Classification with current model
• Create training set from misclassified cliques
• Train regressor/classifier ht
• Augment model
D
19
Illustration Create training set
D
+1
-1
+1
-1
20
Illustration Train regressor ht
h1(∙)
ϕ(∙) = α1 h1(∙)
21
Illustration Classification with current CRF model
ϕ(∙) = α1 h1(∙)
22
Illustration Create training set
ϕ(∙) = α1 h1(∙)
D
-1
+1
23
Illustration Train regressor ht
h2(∙)
ϕ(∙) = α1 h1(∙)+α2 h2(∙)
24
Illustration Stop
ϕ(∙) = α1 h1(∙)+α2 h2(∙)
25
Boosted CRF Related Work Gradient Tree Boosting for CRFs
• Dietterich et al. 2004 Boosted Random Fields
• Torralba et al. 2004 Virtual Evidence Boosting for CRFs
• Liao et al. 2007
Benefits of Max-Margin objective• Do not need marginal probabilities• (Robust) High-order interactions
Kohli et al. 2007, 2008
26
Using Higher Order Information
Colored by elevation
27
Region Based Model
28
Region Based Model
29
Region Based Model
Inference: graph-cut procedure• Pn Potts model (Kohli et al. 2007)
30
How To Train The Model
1Learning
31
How To Train The Model
Learning(ignores features from clique c)
32
How To Train The Model
β
Learning
β1
Robust Pn PottsKohli et al. 2008
33
Experimental Analysis 3-D Point Cloud Classification Geometry Surface Estimation
34
Random Field Description Nodes: 3-D points Edges: 5-Nearest Neighbors Cliques: Two K-means segmentations
Features[0,0,1]
normalθ
Local shape Orientation Elevation
35
Qualitative Comparisons
Parametric Functional (this work)
36
Qualitative Comparisons
Parametric Functional (this work)
37
Qualitative Comparisons
Parametric Functional (this work)
38
Quantitative Results (1.2 M pts) Macro* AP: Parametric 64.3% Functional 71.5%
Wire Pole/Trunk Façade Vegetation0.75
0.85
0.95
0.90
0.80
0.89 0.880.90
0.81
0.88
0.93
Recall0.00
0.20
0.40
0.60
0.80
1.00
0.26 0.22
0.880.99
0.50
0.26
0.910.99
Precision
+24%
+5%
+4%
+3%
-1%
39
Experimental Analysis 3-D Point Cloud Classification Geometry Surface Estimation
40
Random Field Description Nodes: Superpixels (Hoiem et al. 2007) Edges: (none) Cliques: 15 segmentations (Hoiem et al. 2007)
Features (Hoiem et al. 2007)• Perspective, color, texture, etc.
1,000 dimensional space
More Robust
41
Quantitative Comparisons
Parametric (Potts)
Functional (Potts)
Hoiem et al. 2007
Functional (Robust Potts)
42
Qualitative Comparisons
Parametric (Potts)
Functional (Potts)
43
Qualitative Comparisons
Parametric (Potts)
Functional (Potts) Functional (Robust Potts)
44
Qualitative Comparisons
Parametric (Potts)
Functional (Potts)
Hoiem et al. 2007
Functional (Robust Potts)
45
Qualitative Comparisons
Parametric (Potts)
Functional (Potts)
Hoiem et al. 2007
Functional (Robust Potts)
46
Qualitative Comparisons
Parametric (Potts)
Functional (Potts)
Hoiem et al. 2007
Functional (Robust Potts)
47
Conclusion Effective max-margin learning of high-order CRFs
• Especially for large dimensional spaces• Robust Potts interactions• Easy to implement
Future work• Non-linear potentials (decision tree/random forest)• New inference procedures:
Komodakisand Paragios 2009Ishikawa 2009Gould et al. 2009Rother et al. 2009
48
Thank you Acknowledgements
• U. S. Army Research Laboratory• Siebel Scholars Foundation• S. K. Divalla, N. Ratliff, B. Becker
Questions?